Feasibility of a Large Language Model Chatbot to Support Parental Understanding in the PICU

Area of Science:

• Artificial Intelligence in Medicine
• Clinical Informatics
• Patient Engagement Technologies

Background:

• Parental understanding of critical care is vital for family-centered care.
• Large language models (LLMs) offer potential for patient education and support.
• Existing communication tools in the Pediatric Intensive Care Unit (PICU) may not fully meet parental information needs.

Purpose of the Study:

• To assess the feasibility of an LLM-based chatbot for answering parental questions in the PICU.
• To gather data informing the design of a future randomized controlled trial (RCT).

Main Methods:

• A prospective, single-arm feasibility study involving 14 parents of children in a quaternary PICU.
• Parents interacted with a HIPAA-compliant GPT-4o chatbot for 10-minute sessions, using patient-specific electronic health record (EHR) data.
• Feasibility assessed via parental engagement, satisfaction, provider perceptions, accuracy, safety, and recruitment.

Main Results:

• High recruitment rate (87.5%) and parental engagement (median 6 questions).
• Excellent real-time satisfaction (96%) and perceived value (Net Promoter Score [NPS] +57).
• Chatbot responses were highly accurate (99.3%), with minor errors; providers rated quality highly (median 5.0/6.0).

Conclusions:

• An EHR-informed LLM chatbot is feasible for use in the PICU.
• The technology demonstrated high parental engagement, satisfaction, and provider acceptance.
• Findings support the progression to a larger randomized controlled trial (RCT).